Towards a Conversation Analysis Approach for Multicommunicating in Open-Domain, Retrieval-Based Conversational Agents
Monday, May 18, 2020, 10:00am - 12:00pm
Speaker: Carlos M. Muniz
Location : Remote via Webex
Dr. Mubbasir Kapadia, Dr. Gerard de Melo, Dr. Karl Stratos, and Dr. Sepehr Assadi (External Member)
Event Type: Qualifying Exam
Abstract: Multicommunication is a prevalent behavior found in online human conversation, and yet mosthuman-to-chatbot approaches do not address it. Intrinsic parts of human-to-humanconversations like the flexibility of communication tempo, the compartmentalization ofconversation, and the topics and intensity of interactions are elevated in multicommunicativehuman-to-chatbot conversations. Previous Deep Learning approaches may overlook theseintrinsic characteristics due to them focusing on the evaluation of human-likeness throughsensible and specific responses [Google’s Meena] or on-topic responses [Amazon’s Alexa Prize Challenge]. In order to both capture the human-likeness of sensible, specific, and on-topicresponses, we use the fundamental unit of conversation as defined by Conversational Analysis: the adjacency pair; and expand it into an Adjacency Relation to describe a succession of adjacency pairs. To illustrate this Conversational Analysis and Multicommunication Theory, we are developing an Interactive Behavior Tree Conversation Management Framework for Multicommunicating in an Open-Domain, Retrieval-Based Chatbot. We will describe a preliminary deployment of this framework and highlight its advantages (and limitations). We show that this framework provides additional human-likeness not provided by other response retrieval techniques and show this using our own Multi-Engagement Metric. Furthermore, we describe a real problem where a solution of this type will be applied and studied.